A comparative study of multi-scale image super-resolution techniques

Abstract

Super-resolution imaging is a technique that can be used to construct high-resolution imagery from low-resolution images. Low resolution images are often of the same scene, but contain aliasing and different subpixel shifts, which when combined can increase high frequency components while removing blurring. Superresolution reconstruction techniques include methods such as the Interpolation Approach (IA) [1] or the Frequency Domain Approach (FDA) [2]. The IA and FDA methods make use of aliasing in low-resolution images as well as the shifting property of the Fourier transform. Problems arise with both approaches, such as use of blurred images for creating non-optimal reconstructions. In this paper, we propose a comparative analysis if IA methods using reference and no-reference evaluation metrics. The IA methods used are Sparse Reconstruction and Example-based Learning Reconstruction. We evaluate the performance of each algorithms with Root Mean Square Error (RMSE), Peak- Signal-Noise Ratio (PSNR), Mutual Information (MI), Tsallis Entropy (TE), Non-linear Correlation (NLC), Multi-Scale Decomposition (MSD), Spatial Frequency (SF), Structural Similarity Index (SSIM), and Piella’s Metric. Initial results are show that Sparse Reconstruction performs better when evaluating with the proposed metrics, which indicate that this non-linear based interpolation techniques have good predictive capabilities.